Three-dimensional metallogenic prediction method based on full incidence relation learning

Abstract

The embodiment of the invention provides a three-dimensional metallogenic prediction method based on full incidence relation learning, and belongs to the technical field of data processing, and the method specifically comprises the steps: 1, fusing feature engineering and multi-dimensional information through data driving, and obtaining an ore control factor matrix; step 2, constructing monomer energy expression according to the ore control factor matrix; step 3, constructing an interaction energy mechanism of space and feature coupling according to an association relationship between voxels of the mineralization space; 4, constructing a global prediction framework based on monomer energy expression and an interactive energy mechanism; step 5, performing end-to-end-based multi-level parameter optimization on the global prediction framework, optimizing edge distribution of each voxel label and forming target distribution; and step 6, predicting the spatial distribution rule of the mineralization space according to the target distribution. Through the scheme of the invention, the prediction efficiency, accuracy and reliability are improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data processing technology, and in particular to a three-dimensional mineralization prediction method based on full association relationship learning. Background Art

[0002] At present, although machine learning has made certain progress in the application of mineral resource prediction, there are still many shortcomings in the prediction of deep ore bodies. First, most machine learning models rely on the assumption of independent and identical distribution, while mineral resource data has significant spatial correlation, and there are complex geological structures and spatial continuity between elements. Ignoring these correlations will cause the model to be unable to fully capture the mineralization laws of ore deposits. Secondly, the existing models mainly focus on point data or local regional characteristics, and cannot effectively model the global pattern in the three-dimensional structure of the ore deposit, and lack a comprehensive understanding of the spatial variability, local anomalies and the relationship between mineralized elements. In addition, the data types involved in mineral resource exploration are complex and diverse, including geological structure, mineralization distribution, exploration records, geophysics, geochemistry and other information. The existing machine learning models have limited ability to integrate these multi-source heterogeneous data, and it is difficult to extract high-dimensional and potential mineralization characteristics from them. More importantly, the existing methods focus on the prediction of shallow ore bodies, and often lack sufficient accuracy and reliability when it comes to the positioning of deep hidden ore bodies.

[0003] It can be seen that there is an urgent need for a three-dimensional mineralization prediction method based on full correlation learning with high prediction efficiency, accuracy and reliability. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a three-dimensional mineralization prediction method based on full association relationship learning, which at least partially solves the problems of poor prediction efficiency, accuracy and reliability in the prior art.

[0005] The embodiment of the present invention provides a three-dimensional mineralization prediction method based on full association relationship learning, comprising: Step 1: Obtain the mineral control factor matrix through data-driven fusion of feature engineering and multi-dimensional information; Step 2, construct the monomer energy expression according to the mineral control factor matrix; Step 3, construct the interactive energy mechanism of space and feature coupling according to the correlation between each voxel in the mineralized space; Step 4: construct a global prediction framework based on monomer energy expression and interaction energy mechanism; Step 5: Perform end-to-end multi-level parameter optimization on the global prediction framework, optimize the edge distribution of each voxel label and form a target distribution; Step 6: predict the spatial distribution pattern of mineralization space based on the target distribution.

[0006] According to a specific implementation of the embodiment of the present invention, step 1 specifically includes: Step 1.1, extracting ore control information from different aspects; Step 1.2, using the preset method, gradually quantify the control mechanism of the ore-controlling factors in the ore-controlling information on the spatial distribution of the ore body to form an ore-controlling factor matrix.

[0007] According to a specific implementation of the embodiment of the present invention, before step 2, the method further includes: The mineralization control factors corresponding to the sample voxel data set are input into the spatial pattern learning network to obtain the predicted mineralization probability corresponding to each sample voxel; The predicted mineralization probability corresponding to each sample voxel is compared with its corresponding true label to obtain the single-body loss function and optimize the spatial pattern learning network accordingly.

[0008] According to a specific implementation of the embodiment of the present invention, step 2 specifically includes: Step 2.1: Based on the ore-controlling factors and the spatial distribution of the ore body, a spatial pattern learning network is used to construct a single term to represent the energy state of each element in the mineralization space, and the predicted mineralization probability of each element is obtained. in, is the flattened eigenvector, and are the weight matrix and bias respectively, is a given known ore-controlling factor Under the conditions Individual element belongs to predicted mineralization probability of the class; Step 2.2, take the negative logarithm of the predicted mineralization probability to obtain the monomer energy expression in, Represents the energy value of a single cell. represents the probability of mineralization, represents the ore-controlling factor, Indicates Individual element in From the categories, we get category 1.

[0009] According to a specific implementation of the embodiment of the present invention, step 3 specifically includes: Step 3.1, define the label compatibility function between each pair of voxels ; in, Representation labels are usually discretized into a finite set of states A state in Step 3.2: Calculate the spatial correlation between voxels i and j using the distance function in, and It is a voxel and The characteristic vector of and are the spatial three-dimensional coordinates of voxel i and voxel j, and is a constant weight used to balance the effects of differences in ore-controlling factors and spatial distance; Step 3.3: Use the spatial pattern learning network to extract multi-scale information of features layer by layer, and enhance the feature vector through nonlinear transformation and Expressive ability, obtain multi-scale feature information; Step 3.4: Use the contrast loss to train the spatial pattern learning network to obtain the feature similarity function , and the interaction weight is calculated accordingly , where the expression of the contrast loss is in, represents the total number of voxel pairs, It is a voxel and If the label relationship between and Belong to the same category , otherwise ; is the feature similarity function, extracted by the spatial pattern learning network Layer feature representation Calculated, is the target distance between feature similarity pairs, indicating the expected distance between feature vectors corresponding to similar voxels, is the minimum distance threshold of feature non-similar pairs, used to separate the feature space of non-similar pairs, Represents the Euclidean norm of a vector; The interaction weight The expression is in, is the feature similarity function; Step 3.5: Fusion of label compatibility function, spatial correlation, multi-scale feature information and interaction weights , we get the interaction energy mechanism in, Representation voxel and The interaction potential between is the label compatibility function, which ensures that the potential energy is low when the labels of adjacent voxels are the same. is the interaction weight after dynamic optimization between voxels, which adjusts the label dependency between adjacent voxels. It is a voxel and The spatial correlation between the differences in ore-controlling factors and three-dimensional spatial positions is integrated.

[0010] According to a specific implementation of an embodiment of the present invention, the expression of the global prediction framework is: in, Indicates Individual element labels The corresponding monomer potential energy, Indicates Individual element and The interaction potential between individual elements, Represents the total number of voxels.

[0011] According to a specific implementation of the embodiment of the present invention, step 5 specifically includes: Step 5.1: Calculate the conditional probability distribution of the mineralization space according to the global prediction framework in, is the normalization factor; Step 5.2, convert the conditional probability distribution problem into the target distribution problem, and update and optimize the marginal distribution of each voxel label accordingly, where the expression of the update optimization is: ; in, represents the independent marginal distribution of voxel labels; voxel The range is between, represents the total number of voxels; Step 5.3: After the independent edge distribution of all voxel labels is updated and optimized, multiply them to obtain the target distribution .

[0012] The three-dimensional mineralization prediction scheme based on full-association learning in the embodiment of the present invention includes: step 1, obtaining the mineralization control factor matrix through data-driven fusion of feature engineering and multidimensional information; step 2, constructing a single energy expression according to the mineralization control factor matrix; step 3, constructing an interactive energy mechanism of space and feature coupling according to the correlation relationship between each voxel in the mineralization space; step 4, constructing a global prediction framework based on the single energy expression and the interactive energy mechanism; step 5, performing end-to-end multi-level parameter optimization on the global prediction framework, optimizing the edge distribution of each voxel label and forming a target distribution; step 6, predicting the spatial distribution law of the mineralization space according to the target distribution.

[0013] The beneficial effects of the embodiments of the present invention are as follows: through the scheme of the present invention, the complex spatial correlation in the mineralized space is fully considered, and the accurate positioning and prediction of the mineralized body elements are achieved through the end-to-end deep learning architecture. This method can effectively integrate local information with global information during the model training process, greatly improving the prediction accuracy and reliability of deep ore bodies. By making full use of the interaction between mineralization probability and interaction potential energy in the mineralized space, a more refined prediction of mineralized bodies is achieved, providing a new intelligent modeling method for three-dimensional mineralization prediction, and improving prediction efficiency, accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 A schematic flow chart of a three-dimensional mineralization prediction method based on full association relationship learning provided by an embodiment of the present invention; Figure 2 A model effect comparison diagram provided by an embodiment of the present invention; Figure 3 A case result diagram provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0017] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0018] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present invention, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0019] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention. The drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0020] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0021] An embodiment of the present invention provides a three-dimensional mineralization prediction method based on full association relationship learning, which can be applied to the mineralization prediction process in mining scenarios.

[0022] See also Figure 1 , which is a flow chart of a three-dimensional mineralization prediction method based on full correlation learning provided by an embodiment of the present invention. Figure 1 As shown, the method mainly comprises the following steps: Step 1: Obtain the mineral control factor matrix through data-driven fusion of feature engineering and multi-dimensional information; In the specific implementation, we can take the ore-controlling law as the core and construct a complete and systematic feature system through multi-dimensional and multi-scale information extraction. Specifically, we will comprehensively extract ore-controlling information from aspects such as spatial expansion effect, fluid migration path, fluid diffusion behavior and implicit morphological characteristics. We use advanced technical methods such as Markov chain simulation, Darcy fluid kinetic energy conversion and Laplace characteristic function to gradually quantify the control mechanism of ore-controlling factors on the spatial distribution of ore bodies. In addition, combined with the specific geological background of the ore deposit, the existing mineralization information indicators are reorganized and optimized to ensure that the extracted features can accurately reflect the ore-controlling law, form a systematic and targeted mineralization information extraction plan, and lay a solid data foundation for subsequent modeling and prediction.

[0023] Step 2, construct the monomer energy expression according to the mineral control factor matrix; When implementing it, it is necessary to obtain the mining control indicators Mineralization distribution The relationship between and the calculation of monomer potential energy includes the following steps: S21. Establish the basic model of single items; S22, calculate monomer potential energy; In S21, we use the Spatial Pattern Learning Network (SPLN) to describe the ore control indicators. Based on the relationship between the mineralization control factors and prospecting information, we construct monomer terms to represent the energy state of a single voxel in the mineralization space. The core task of the monomer term is to quantitatively evaluate the mineralization potential of each voxel, which directly affects the prediction effect of the overall model. In the implementation process, we use SPLN as the basic model. Through its powerful feature extraction and nonlinear mapping capabilities, we extract high-level features from the mineralization control factors and prospecting information and map them to the mineralization probability of the voxel. Specifically, we first input the mineralization control factors into the SPLN model. Through convolution operations, the network extracts correlation information in spatial and feature dimensions layer by layer, thereby capturing the potential laws in the mineralization space. At the same time, the pooling layer in the network further enhances the ability to integrate mineralization information at different scales, making the model more adaptable to complex spatial distribution patterns. The output layer adopts a probabilistic design and finally outputs the mineralization probability of each voxel as its corresponding monomer term value. These probabilities not only reflect the mineralization potential of each voxel, but also ensure consistency with the overall prediction target through the end-to-end training process of the network. Compared with traditional linear models or artificially designed feature methods, SPLN can more comprehensively capture the nonlinear relationship between ore-controlling factors and prospecting information, thereby significantly improving the accuracy of the evaluation of the mineralization potential of the volume element. This innovative design ensures the efficiency and accuracy of the single item, and provides a reliable foundation for the construction of subsequent interaction items and the optimization of energy functions.

[0024] To specify the construction process of the single item, we built an end-to-end spatial pattern learning network (SPLN) to quantify the mineralization potential of a single element. The input of the mineralization factor is expressed in matrix form as , where N represents the number of voxels and M is the dimension of the mineralization control factor, such as structure, temperature field, distance field and other indicators. For each voxel i, its mineralization control factor can be expressed as a vector We use SPLN to extract the spatial relationship and high-order correlation among the mineral control factors. The entire network structure can be described as: (1) Input layer: Input the mineral control factor matrix X into the network as the initial feature representation; (2) Convolutional layer: Extract feature patterns through multi-layer convolution operations: ; in It is The feature map of the layer, and denote the convolution kernel and bias respectively, represents the convolution operation, is the activation function (such as ReLU).

[0025] (3) Pooling layer: Pooling operation is used to further compress feature dimensions and enhance the scale invariance of the model: ; (4) Fully connected layer: The outputs of the convolutional layer and the pooling layer are flattened and mapped to the voxel mineralization probability space through the fully connected layer: ; in is the flattened eigenvector, and are the weight matrix and bias respectively, ∈[0, 1] is a given known mineral control factor Under the conditions Individual element belongs to The mineralization probability of the category.

[0026] (5) Loss function design: In order to optimize the SPLN model, we use a single loss function suitable for our research task to predict the mineralization probability With the true label For comparison, the formula is: ; in, , represents the true label, Represents the probability of mineralization predicted by the model.

[0027] (6) End-to-end training: Through the above structure, input the control factor , directly output the mineralization probability of each voxel , and by optimizing the loss function , spatial features and mineralization patterns are learned simultaneously during the training process, thus ensuring the efficiency and accuracy of individual items.

[0028] Furthermore, in S22, we have calculated the probability of mineralization of each voxel, and we take the negative logarithm to obtain the monomer energy, which is calculated as follows: ; in, Represents the energy value of a single cell. represents the probability of mineralization, represents the ore-controlling factor, Indicates Individual element in From the categories, we get category 1.

[0029] Step 3, construct the interactive energy mechanism of space and feature coupling according to the correlation between each voxel in the mineralized space; In the specific implementation, this step needs to consider the correlation between mineralized voxel i and mineralized voxel j and calculate it as interaction potential. These correlations are not only related to their label similarity, but also to their spatial coordinates and feature similarity, including the following steps: S31, label association modeling; S32, spatial geometry dependency analysis; S33, feature multi-scale aggregation; S34, dynamic weight fusion; S35. Calculation of interaction potential energy.

[0030] In S31, the label association rules between voxels are established to characterize their compatibility. We first define the label context penalty term between each pair of voxels To quantify the compatibility relationship between mineralized voxel i and mineralized voxel j under different label configurations. The key to this step is to incorporate the semantic compatibility between labels into potential energy modeling. For example, if voxel and If the label configuration is highly relevant in geological terms, then its label penalty value is low. Through this modeling, the contextual adaptability of the model to mineralization labels is enhanced. The formula is as follows: ; in, represents the label penalty term between voxels i and j, which is determined by the learned model and represents the interaction between labels. For example, if label u and label v have a strong contextual relationship in space, then will be smaller; on the contrary, if they are located in incompatible spatial coordinates, Will be larger. and are the labels of voxels i and j in categories u and v respectively.

[0031] Since the interaction term depends not only on whether the labels of adjacent voxels are the same, but also on their feature similarity and coordinate differences, in S32, to characterize the geometric structure features and spatial position dependence between voxels in three-dimensional space, we calculate the spatial correlation between voxels i and j, which is partly calculated by the distance function To express it. Common distance functions include the Euclidean distance of voxels, combined with factors such as ore control factors and coordinates. The formula is as follows: ; in, and are the eigenvectors of voxels i and j, and are the three-dimensional coordinates of voxel i and voxel j. and is a constant weight used to balance the influence of differences in ore-controlling factors and spatial distance. This modeling process integrates the spatial distribution patterns of ore-controlling factors and geometric structures, ensuring that the model accurately depicts the three-dimensional spatial dependence.

[0032] In the mineralization space, the correlation between features may have multi-level characteristics. In order to capture the interaction pattern of different scale information in the mineralization space through the multi-scale feature extraction mechanism, in S33, we use SPLN to extract the multi-scale information of features layer by layer and enhance the features through nonlinear transformation. and Expressive ability:

[0033] in It is the volume element i The feature map of the layer, and denote the convolution kernel and bias respectively, represents the convolution operation, is the activation function (such as ReLU), and the voxel j parameter is the same. This part integrates the mineralization features of different scales and provides context feature support for the subsequent interaction potential energy calculation.

[0034] In S34, in order to assign dynamic weights to the interactions between voxels to adapt to the changes in their labels, positions, and feature relationships, we introduce learnable parameters , according to the feature similarity and spatial position relationship between voxels, the interaction strength is adaptively adjusted: ; in, is a feature similarity function, which is obtained by training the SPLN network using contrast loss. This step dynamically optimizes the interaction strength, making the model more adaptable to complex spatial distribution patterns and laying a weight foundation for the reasonable allocation of interaction potential energy. Its formula is: ; in, represents the total number of voxel pairs, It is a voxel and If the label relationship between and Belong to the same category , otherwise ; is the feature similarity function, extracted by SPLN Layer feature representation Calculated; is the target distance of feature similarity pairs, which represents the expected distance between feature vectors corresponding to similar voxels; is the minimum distance threshold of feature non-similar pairs, used to separate the feature space of non-similar pairs. Represents the Euclidean norm of a vector.

[0035] In S35, in order to integrate label compatibility, spatial geometric dependence and multi-scale feature information, an efficient interaction potential energy function is constructed. We integrate S31-S34 to obtain the final interaction potential energy calculation formula: ; in, Representation voxel and The interaction potential between is a label compatibility function that ensures that the potential energy is low when adjacent voxels have the same label. It is the interaction weight after dynamic optimization between voxels, which regulates the label dependency between adjacent voxels. It is a voxel and It integrates the differences in ore-controlling factors and the dependence on three-dimensional spatial position.

[0036] Step 4: construct a global prediction framework based on monomer energy expression and interaction energy mechanism; In specific implementation, in this step, we hope to associate the model's objectives with the input features by constructing a total energy function, and obtain the optimal label configuration by minimizing the energy function. Therefore, we need to aggregate the monomer potential energy and the interaction potential energy to construct a total energy function, which can be expressed as the sum of the monomer term and the interaction term: ; in, represents the total energy; Indicates Individual element labels The corresponding monomer potential energy, Indicates Individual element and The interaction potential between individual elements; Represents the total number of voxels.

[0037] Step 5: Perform end-to-end multi-level parameter optimization on the global prediction framework, optimize the edge distribution of each voxel label and form a target distribution; In specific implementation, the model can be trained by minimizing the total energy function and optimized by the standard back propagation algorithm. The conditional probability distribution can be calculated by the following formula: ; in In a given feature The conditional probability of label configuration y in the case of . is a normalization factor to ensure is a valid probability distribution.

[0038] To optimize label prediction, our goal is to find the optimal label configuration , so that the total energy function Minimize, which is equivalent to maximizing the conditional probability P(Y=y|X) of the label configuration, that is: ; However, directly solving It is necessary to calculate all possible label configurations Y, and the computational complexity is exponential, so we use approximate inference to speed up the calculation. In approximate inference, we use a simpler distribution To approximate the true distribution , this distribution It can be decomposed into the product of independent marginal distributions of each voxel label: ; In order to approximate the distribution As close to the true distribution as possible , we minimize the Kullback-Leibler (KL) divergence between them to optimize: ; in, is an approximate distribution; is the true distribution. After the loss of the total energy function is obtained, we combine formulas (4), (11) and (17) to obtain the final total loss function: ; in, and is a hyperparameter. By minimizing the total loss function through the standard back-propagation algorithm, we achieve end-to-end joint optimization of unary potential energy, interaction potential energy, and global energy, thereby improving the performance of the model in the label prediction task.

[0039] By distributing the edge By optimizing, we can get the approximate marginal distribution of each voxel label. Specifically, for each i, we Update and solve the following update formula: ; in, is represented as a normalized distribution, In addition to The expectation of other variables besides , is expanded to: ; For the monomer potential function , which depends on , does not involve the expectations of other variables, so, , for the interaction term, , so, expand the expectation term to get: ; in, Representation and The adjacent node set. Finally, the update formula is: ; This update formula means that we iterate and calculate each , to gradually optimize the probability distribution of label predictions. The specific algorithm is shown in Table 1. Finally, after a certain number of iterations, we can use these approximate marginal distributions to calculate the conditional probability of label configurations. The whole process includes the construction of the total energy function, the calculation of monomer and interaction potential energy, the minimization of the KL divergence of approximate inference, and the optimization of the model by iteratively updating the marginal distribution. These steps work together to obtain the optimal label configuration by minimizing the energy function. The calculation process can be effectively accelerated through approximate inference methods, ultimately achieving intelligent inference and accurate prediction of spatial distribution patterns.

[0040] Table 1

[0041] Step 6: predict the spatial distribution pattern of mineralization space based on the target distribution.

[0042] In specific implementation, after the independent edge distribution of all voxel labels is updated and optimized, they are multiplied to obtain the target distribution. The calculation process can be effectively accelerated through approximate inference methods, and ultimately the intelligent inference and precise prediction of spatial distribution rules can be achieved.

[0043] The three-dimensional mineralization prediction method based on full-association relationship learning provided in this embodiment can play a unique advantage in deep ore body prediction by introducing a new spatial association modeling framework on the basis of traditional machine learning models. First, this method abandons the independent and identically distributed assumption and models the spatial correlation in the deep part of the ore deposit. Our model can effectively capture the interaction between ore bodies, use spatial adjacency information to describe the complex association between voxels, and improve the model's understanding and prediction ability of the ore body distribution law. Secondly, we combine the feature extraction ability of deep learning in the design of model monomer terms and interaction terms. In addition, in order to effectively process the three-dimensional spatial information of the ore deposit, we introduce the coordinate kernel function so that each voxel can establish a connection with its surrounding environment and adjacent voxels, thereby performing global structural modeling. This innovative method can break through the limitations of traditional machine learning models, not only can it achieve accurate positioning of deep hidden ore bodies, but also can provide a more comprehensive and reliable basis for the prediction of deep mineral resources. In practice, this method has proven its great potential in the sustainable development of mineral resources, especially in resource succession exploration in crisis mines and old mines. It provides effective technical support and improves prediction efficiency, accuracy and reliability.

[0044] The method of the present invention will be further described in conjunction with a specific example. A gold mine in a certain place is used as the research object to verify the applicability of the method. In this case, the method is used to achieve a three-dimensional prediction of the distribution of hidden ore bodies. In the experiment, some mainstream machine learning models are selected for comparison. Figure 2 The AUC value of the ROC curve in Figure 2 shows that this method significantly improves the prediction accuracy. In addition, we also calculated the success curve to evaluate the performance of our method. The specific results are as follows: Figure 3 The embodiment of the present invention specifically includes the following steps: S1. Establish monomer potential function; In this study, considering the complexity of mineralization space and the interaction of high-dimensional features, we chose SPLN as the method to establish the monomer potential function. Specifically, we used the SPLN model to estimate the mineralization probability of each element and then quantify the mineralization possibility of each element.

[0045] First, based on the geological background and mineralization concept model of the area, we extracted 42 mineralization dynamic indicators as ore-controlling factors, including strike expansion, dip expansion, fluid channel migration length, channel flow, alteration zone thickness, fracture surface distance field, dip shear fracture favorableness and Laplace characteristic function, etc. These characteristics are used to construct the basic data set required for model training.

[0046] In order to apply SPLN for training, these mineralization control factors are first passed into SPLN as input. The network structure of SPLN consists of multiple convolutional layers, pooling layers, fully connected layers and the final Softmax output layer. The mineralization probability prediction of each voxel is obtained through the output of the last layer of the network.

[0047] The training process of the SPLN model uses the Adam optimizer (learning rate is 1e-5) and the monomer loss function for training (Equation 13). Specifically, the SPLN model is trained with the training set, and the parameters of the convolution kernel and the fully connected layer are updated by the back propagation algorithm during the optimization process to minimize the prediction error on the training data set. Finally, we obtain the probability of each voxel being marked as having or not having a mineral, and this probability value is used as the input of the monomer potential function to quantify the mineralization possibility of the voxel. In this way, through the powerful feature extraction and pattern learning capabilities of SPLN, the mineralization probability of mineralized voxels can be predicted more accurately, providing reliable probability information for the subsequent calculation of interaction terms.

[0048] S2, establish interaction terms; Similar to the monomer potential function, the input data of the interaction term also includes the mineralization space control factors, such as fluid divergence, shear strain, volume strain, etc. These features form a description of the relationship between each pair of voxels by extracting the local and global information of each voxel. In addition, the interaction term needs to consider the spatial coordinate relationship between voxels, so for each pair of adjacent voxels and , we use its features as input to calculate the interaction terms.

[0049] First, using the geological background and exploration data of the area, the ore-controlling factors of each voxel are extracted, including strike expansion, dip expansion, fluid channel migration length, channel flow, alteration zone thickness, fracture surface distance field, dip shear fracture favorableness and Laplace characteristic function channel indicators. In addition, the spatial coordinates of each voxel are are also included in the dataset.

[0050] For each pair of voxels i and j, we first calculate their feature differences. The feature differences here are quantified by calculating the Euclidean distance of the voxel feature vectors, as follows: ; in, and is the eigenvector of voxels i and j. This difference quantifies the difference in the mineral control factor between voxels.

[0051] Next, we calculate the difference in spatial coordinates between voxels i and j. This difference is quantified by calculating the Euclidean distance between voxel coordinates, as follows: ; in,( ) and( ) are the spatial coordinates of voxels i and j. The spatial difference reflects the coordinate relationship of voxels in three-dimensional space, and the smaller the distance, the closer the space.

[0052] According to the feature differences and spatial coordinate differences, we construct the interaction potential energy function, the formula is: ; In this formula: and is a hyperparameter used to adjust the importance of feature differences and spatial coordinate differences in the interaction term. and is the mining factor vector of voxel i and j. ( )and( ) are the spatial coordinates of voxels i and j.

[0053] Through this formula, we combine the feature similarity and spatial relationship between voxels to obtain the interaction potential , which is used to quantify the dependencies between pairs of voxels.

[0054] To combine the interaction potential with the rest of the model, we use SPLN to learn and optimize the interaction potential. SPLN extracts the features of the voxels and constructs a similarity matrix (i.e., interaction potential) based on them, which effectively captures the spatial and feature similarities between voxels.

[0055] SPLN settings: Convolutional layer: SPLN is used to extract high-level features of voxels, which play an important role in calculating interaction potential. Fully connected layer: The fully connected layer is used to map features to the output interaction potential to capture the similarity relationship between voxels. Input features: The input of SPLN is the feature vector and spatial coordinates of each voxel. Training and optimization: The model is trained using the standard back-propagation algorithm. The loss function is shown in Equation (11). Through iterative optimization, the weight of the interaction potential is updated. and , and the convolutional layer weights of the network.

[0056] S3, establishing a total energy function; After calculating the potential energy, the algorithm uses the monomer potential energy and the interaction potential energy to calculate the approximate posterior distribution Logarithm of : ; This means that the energy of the entire system is constructed by the weighted sum of the individual potential energies and the interaction potential energies.

[0057] S4, quantitative prediction of concealed ore bodies taking into account the spatial correlation of mineralization: We hope to associate the model's objectives with the input features by constructing a total energy function, and obtain the optimal label configuration by minimizing the energy function. Therefore, we need to aggregate the monomer potential energy and the interaction potential energy to construct a total energy function, which can be expressed as the sum of the monomer term and the interaction term: ; in, represents the total energy; Indicates Individual element labels The corresponding monomer potential energy, Indicates Individual element and The interaction potential between individual elements; Represents the total number of voxels.

[0058] Furthermore, in S5, we train the model by minimizing the total energy function and optimize it through the standard back-propagation algorithm. The conditional probability distribution can be calculated by the following formula: ; in In a given feature The conditional probability of label configuration y in the case of . is a normalization factor to ensure is a valid probability distribution.

[0059] To optimize label prediction, our goal is to find the optimal label configuration , so that the total energy function Minimize, which is equivalent to maximizing the conditional probability P(Y=y|X) of the label configuration, that is: ; However, directly solving It is necessary to calculate all possible label configurations Y, and the computational complexity is exponential, so we use approximate inference to speed up the calculation. In approximate inference, we use a simpler distribution To approximate the true distribution , this distribution It can be decomposed into the product of independent marginal distributions of each voxel label: ; In order to approximate the distribution As close to the true distribution as possible , we minimize the Kullback-Leibler (KL) divergence between them to optimize: ; in, is an approximate distribution; is the true distribution. After the loss of the total energy function is obtained, we combine formulas (4), (11) and (17) to obtain the final total loss function: ; in, and is a hyperparameter. By minimizing the total loss function through the standard back-propagation algorithm, we achieve end-to-end joint optimization of unary potential energy, interaction potential energy, and global energy, thereby improving the performance of the model in the label prediction task.

[0060] Furthermore, S6 distributes By optimizing, we can get the approximate marginal distribution of each voxel label. Specifically, for each i, we Update and solve the following update formula: ; in, is represented as a normalized distribution, In addition to The expectation of other variables besides , is expanded to: ; For the monomer potential function , which depends on , does not involve the expectations of other variables, so, , for the interaction term, , so expanding the expected term yields: ; in, represents the independent marginal distribution of voxel labels; Representation voxel The range is , Represents the total number of voxels. Finally, the update formula is: ; This update formula means that we iterate and calculate each , to gradually optimize the probability distribution of label predictions. Finally, after a certain number of iterations, we can use these approximate marginal distributions to calculate the conditional probability of label configurations. The whole process includes the construction of the total energy function, the calculation of monomer and interaction potential energy, the minimization of the KL divergence of approximate inference, and the optimization of the model by iteratively updating the marginal distribution. These steps work together to obtain the optimal label configuration by minimizing the energy function The calculation process can be effectively accelerated through approximate inference methods, ultimately achieving intelligent inference and accurate prediction of spatial distribution patterns.

[0061] According to the drilling data and spatial distribution information of the mineral belt, a three-dimensional model of mineralized elements was constructed, and 58,239 three-dimensional spatial elements were finally generated. These elements were determined to be mineralized based on the results of chemical analysis and industrial standards (i.e., gold grade greater than 1 g / ton), and then mineralization information was extracted and assigned to each element. The entire three-dimensional area was divided into small cubes with equal intervals at a resolution of 25 meters, and the volume of each cube was 15,625 cubic meters. In order to simulate the spatial distribution of gold grade, the kriging method was used to interpolate the gold content, and the predicted gold content value was assigned to the spatial monomers in the model. After processing, 58,239 three-dimensional monomers with known gold grades were finally obtained, including 27,955 mineralized monomers and 30,284 non-mineralized monomers.

[0062] In the mineralization prediction task, this problem is defined as a binary classification problem. In the data set, the mineralized monomer is assigned the label 1, and the non-mineralized monomer is assigned the label 0. Based on the spatial coordinates, mineralization characteristics and ore-controlling factors of the mineralized monomer, the feature data set required for training is prepared. Subsequently, we divided the data set into a training set and a test set with a ratio of 7:3, and then started to train our model. In order to prove the effectiveness of our model, we used nonlinear logistic regression (LR), support vector machine (SVM), extreme gradient boosting (XG-Boost), and multi-layer perceptron (MLP) to compare the accuracy with the model we proposed (Ours). The results show that the model we proposed is more advantageous, so we constructed a three-dimensional mineralization prediction model (3DMPM) for this mineral belt.

[0063] The model of the present invention is implemented in the PyTorch-1.8 environment and trained using the NVIDIA GeForce GTX1060Ti GPU. During the training process, the inference algorithm of the association layer was iterated 10 times to ensure the optimization effect. In the training stage, we used data with a batch size of 32. To ensure the stability and adaptability of the model, the stochastic gradient descent (SGD) method with a learning rate of 1e-5 was used for optimization, and the training lasted for 60 cycles. In addition, in order to improve the robustness of the model, each architecture was independently trained 5 times using different random seeds to avoid the impact of the randomness of parameter initialization. In the end, the model we successfully trained has the ability to consider the correlation between voxels in mineral prediction. Its performance is as follows Figure 2 , 3 shown.

[0064] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof.

[0065] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A three-dimensional mineralization prediction method based on full correlation learning, characterized in that: include: Step 1: Obtain the mineral control factor matrix through data-driven fusion of feature engineering and multi-dimensional information; Step 2, construct the monomer energy expression according to the mineral control factor matrix; Step 3, construct the interactive energy mechanism of space and feature coupling according to the correlation between each voxel in the mineralized space; Step 4: construct a global prediction framework based on monomer energy expression and interaction energy mechanism; Step 5: Perform end-to-end multi-level parameter optimization on the global prediction framework, optimize the edge distribution of each voxel label and form a target distribution; Step 6: predict the spatial distribution pattern of mineralization space based on the target distribution.

2. The method according to claim 1, characterized in that The step 1 specifically includes: Step 1.1, extracting ore control information from different aspects; Step 1.2, using the preset method, gradually quantify the control mechanism of the ore-controlling factors in the ore-controlling information on the spatial distribution of the ore body to form an ore-controlling factor matrix.

3. The method according to claim 2, characterized in that Before step 2, the method further includes: The mineralization control factors corresponding to the sample voxel data set are input into the spatial pattern learning network to obtain the predicted mineralization probability corresponding to each sample voxel; The predicted mineralization probability corresponding to each sample voxel is compared with its corresponding true label to obtain the single-body loss function and optimize the spatial pattern learning network accordingly.

4. The method according to claim 3, characterized in that The step 2 specifically includes: Step 2.1: Based on the ore-controlling factors and the spatial distribution of the ore body, a spatial pattern learning network is used to construct a single term to represent the energy state of each element in the mineralization space, and the predicted mineralization probability of each element is obtained. ； in, is the flattened eigenvector, and are the weight matrix and bias respectively, ∈[0,1] is a given known mineral control factor Under the conditions Individual element belongs to predicted mineralization probability of the class; Step 2.2, take the negative logarithm of the predicted mineralization probability to obtain the monomer energy expression in, Represents the energy value of a single cell. represents the probability of mineralization, represents the ore-controlling factor, Indicates that the i-th element is From the categories, we get category 1.

5. The method according to claim 4, characterized in that The step 3 specifically includes: Step 3.1, define the label compatibility function between each pair of voxels ; in, Representation labels are usually discretized into a finite set of states A state in Step 3.2: Calculate the spatial correlation between voxels i and j using the distance function ； in, and It is a voxel and The characteristic vector of and are the spatial three-dimensional coordinates of voxel i and voxel j, and is a constant weight used to balance the effects of differences in ore-controlling factors and spatial distance; Step 3.3: Use the spatial pattern learning network to extract multi-scale information of features layer by layer, and enhance the feature vector through nonlinear transformation and Expressive ability, obtain multi-scale feature information; Step 3.4: Use the contrast loss to train the spatial pattern learning network to obtain the feature similarity function , and the interaction weight is calculated accordingly , where the expression of the contrast loss is in, represents the total number of voxel pairs, It is a voxel and If the label relationship between and Belong to the same category , otherwise ; is the feature similarity function, extracted by the spatial pattern learning network Layer feature representation Calculated, is the target distance between feature similarity pairs, indicating the expected distance between feature vectors corresponding to similar voxels, is the minimum distance threshold of feature non-similar pairs, used to separate the feature space of non-similar pairs, Represents the Euclidean norm of a vector; The interaction weight The expression is ； in, is the feature similarity function; Step 3.5: Fusion of label compatibility function, spatial correlation, multi-scale feature information and interaction weights , we get the interaction energy mechanism in, Representation voxel and The interaction potential between is the label compatibility function, which ensures that the potential energy is low when the labels of adjacent voxels are the same. is the interaction weight after dynamic optimization between voxels, which adjusts the label dependency between adjacent voxels. It is a voxel and The spatial correlation between the differences in ore-controlling factors and three-dimensional spatial positions is integrated.

6. The method according to claim 5, characterized in that The expression of the global prediction framework is in, Indicates Individual element labels The corresponding monomer potential energy, Indicates Individual element and The interaction potential between individual elements, Represents the total number of voxels.

7. The method according to claim 6, characterized in that The step 5 specifically includes: Step 5.1: Calculate the conditional probability distribution of the mineralization space according to the global prediction framework in, is the normalization factor; Step 5.2, convert the conditional probability distribution problem into the target distribution problem, and update and optimize the marginal distribution of each voxel label accordingly, where the expression of the update optimization is: ； in, represents the independent marginal distribution of voxel labels, Representation voxel The range is between, represents the total number of voxels; Step 5.3: After the independent edge distribution of all voxel labels is updated and optimized, multiply them to obtain the target distribution 。

Images